File size: 11,096 Bytes
89682f8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
from contextlib import nullcontext
from io import BytesIO
import os
import threading
from typing import Optional, Union
import warnings

from compel import Compel
from fastapi.responses import StreamingResponse
from loguru import logger
from PIL import Image
import torch

from leptonai.photon import Photon, FileParam, get_file_content, HTTPException


EXAMPLE_IMAGE_BASE64 = "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"


class JPEGResponse(StreamingResponse):
    media_type = "image/jpeg"


class ImgPilot(Photon):
    requirement_dependency = [
        "torch",
        "diffusers",
        "invisible-watermark",
        "compel",
        "Pillow",
    ]

    # In default, we will use gpu.a10 as the computation resource shape. This should
    # be fast enough.
    deployment_template = {
        "resource_shape": "gpu.a10",
        "env": {
            "MODEL": "SimianLuo/LCM_Dreamshaper_v7",
            "USE_TORCH_COMPILE": "false",
            "WIDTH": "768",
            "HEIGHT": "768",
            "PRINT_PROMPT": "false",
        },
    }

    # A10 should be able to support a maximum concurrency of 8 requests to interleave
    # IO and compute. This is not tuned by the way.
    handler_max_concurrency = 1

    def init(self):
        from diffusers import AutoPipelineForImage2Image  # type: ignore

        cuda_available = torch.cuda.is_available()

        if cuda_available:
            self.device = torch.device("cuda")
        else:
            self.device = torch.device("cpu")

        self.base = AutoPipelineForImage2Image.from_pretrained(
            os.environ["MODEL"],
            torch_dtype=torch.float16 if cuda_available else torch.float32,
        )
        self.base.safety_checker = None
        self.base.requires_safety_checker = False
        if self.handler_max_concurrency > 1:
            self.base_lock = threading.Lock()
        else:
            self.base_lock = nullcontext()
        self.print_prompt = os.environ["PRINT_PROMPT"].lower() in [
            "true",
            "t",
            "1",
            "yes",
            "y",
        ]
        logger.info(f"print_prompt: {self.print_prompt}")
        if cuda_available:
            self.base.to("cuda")
            self.use_torch_compile = os.environ["USE_TORCH_COMPILE"].lower() in [
                "true",
                "t",
                "1",
                "yes",
                "y",
            ]
            if self.use_torch_compile:
                if self.handler_max_concurrency > 1:
                    warnings.warn(
                        "torch compile does not support multithreading, so we will"
                        " disable torch compile since handler_max_concurrency > 1."
                    )
                else:
                    self.width = int(os.environ["WIDTH"])
                    self.height = int(os.environ["HEIGHT"])
                    logger.info(
                        "Compiling model with torch.compile. Note that with torch"
                        " compile, your first invocation will be slow, but subsequent"
                        " invocations will be faster."
                    )
                    self.base.unet = torch.compile(
                        self.base.unet, mode="reduce-overhead", fullgraph=True
                    )
        else:
            self.use_torch_compile = False

        self.compel_proc = Compel(
            tokenizer=self.base.tokenizer,
            text_encoder=self.base.text_encoder,
            truncate_long_prompts=False,
        )  # type: ignore

        logger.info(f"Initialized model {os.environ['MODEL']}. cuda: {cuda_available}.")

    @Photon.handler(
        "run",
        example={
            "prompt": (
                "Portrait of The Terminator, glare pose, detailed, intricate, full of"
                " colour, cinematic lighting, trending on artstation, 8k,"
                " hyperrealistic, focused, extreme details, unreal engine 5, cinematic,"
                " masterpiece"
            ),
            "seed": 2159232,
            "strength": 0.5,
            "steps": 4,
            "guidance_scale": 8.0,
            "width": 512,
            "height": 512,
            "lcm_steps": 50,
            "input_image": EXAMPLE_IMAGE_BASE64,
        },
    )
    def run(
        self,
        prompt: str,
        seed: int,
        strength: float,
        steps: int,
        guidance_scale: float,
        width: int,
        height: int,
        lcm_steps: int,
        input_image: Optional[Union[str, FileParam]],
    ) -> JPEGResponse:
        from diffusers.utils import load_image  # type: ignore
        import time

        start = time.time()

        if self.print_prompt:
            logger.info(f"Prompt: {prompt}")

        # diffusers truncates prompt to 77 tokens, in case prompt is too long, we will
        # use compel to process the prompt (but compel is slower)
        tokens = self.base.tokenizer(prompt, return_tensors="pt")
        if tokens.input_ids.shape[1] > 77:
            prompt_embeds = self.compel_proc(prompt)
            prompt = None
        else:
            prompt_embeds = None

        if input_image is not None:
            image_file = get_file_content(input_image, return_file=True)
            pil_image = Image.open(image_file, formats=["JPEG", "PNG", "GIF", "BMP"])
            if self.use_torch_compile:
                # checks width and height parameter, and return error if width and height are not correct
                if width != self.width or height != self.height:
                    raise HTTPException(
                        status_code=400,
                        detail=(
                            f"width and height must be {self.width} and"
                            f" {self.height} when use_torch_compile is true."
                        ),
                    )
                # checks input image height and width, and resize if necessary
                if pil_image.height != self.height or pil_image.width != self.width:
                    pil_image = pil_image.resize(
                        (self.width, self.height), Image.BILINEAR
                    )
            input_image = load_image(pil_image).convert("RGB")

        with self.base_lock:
            generator = torch.manual_seed(seed)
            output_image = self.base(
                prompt=prompt,
                prompt_embeds=prompt_embeds,
                generator=generator,
                image=input_image,
                strength=strength,
                num_inference_steps=steps,
                guidance_scale=guidance_scale,
                width=width,
                height=height,
                lcm_origin_steps=lcm_steps,
                output_type="pil",
            )  # type: ignore

        nsfw_content_detected = (
            output_image.nsfw_content_detected[0]
            if "nsfw_content_detected" in output_image
            else False
        )  # type: ignore
        if nsfw_content_detected:
            raise HTTPException(status_code=400, detail="nsfw content detected")
        else:
            img_io = BytesIO()
            output_image.images[0].save(img_io, format="JPEG")  # type: ignore
            img_io.seek(0)
            logger.info(f"Produced output in {time.time() - start} seconds.")
            return JPEGResponse(img_io)


if __name__ == "__main__":
    p = ImgPilot()
    p.launch()